Understanding the Scheduling Performance in Wireless Networks with ...

1

Understanding the Scheduling Performance in Wireless Networks with Successive Interference Cancellation Shaohe Lv1 , Weihua Zhuang2, Ming Xu1a , Xiaodong Wang1 , Chi Liu1b , and Xingming Zhou1 1 National Laboratory of Parallel and Distributed Processing National University of Defense Technology, Changsha, P. R. China Email: {shaohelv, xdwang, xmzhou}@nudt.edu.cn, a [email protected], b [email protected] 2 Department of Electrical and Computer Engineering University of Waterloo, Waterloo, Ontario, Canada Email: [email protected]

Abstract—Successive interference cancellation (SIC) is an effective way of multipacket reception to combat interference in wireless networks. We focus on link scheduling in wireless networks with SIC, and propose a layered protocol model and a layered physical model to characterize the impact of SIC. In both the interference models, we show that several existing scheduling schemes achieve the same order of approximation ratios, independent of whether or not SIC is available. Moreover, the capacity order in a network with SIC is the same as that without SIC. We then examine the impact of SIC from first principles. In both chain and cell topologies, SIC does improve the throughput with a gain between 20% and 100%. However, unless SIC is properly characterized, any scheduling scheme cannot effectively utilize the new transmission opportunities. The results indicate the challenge of designing an SIC-aware scheduling scheme, and suggest that the approximation ratio is insufficient to measure the scheduling performance when SIC is available. Keywords-Network capacity; link scheduling; successive interference cancellation

I. INTRODUCTION The capacity of a modern wireless communication system is interference-limited. Due to the broadcast nature, what is arriving at a receiver is a composite signal from all near-by transmissions. In general, the receiver tries to decode only one transmission by regarding all the others as interference and noise. When the arrivals of multiple transmissions overlap, collision occurs and the reception fails. Multiple packet reception (MPR) is a promising technique at the physical layer to combat the interference. When the links interfering with each other transmit simultaneously, a receiver node can separate the collided signals with the MPR capability. It is shown in [1–4] that MPR can significantly increase the capacity of a wireless network. SIC is an effective way of MPR to resolve the transmission collisions [5]. With SIC, the receiver tries to detect multiple received signals using an iterative approach. In each iteration, the strongest signal is decoded, by treating the remaining signals as interference. If a required SINR (signal to interference and noise ratio) is satisfied, this signal can be decoded and removed from the received composite signal. In the subsequent

iteration, the next strongest signal is decoded, and the process continues until either all the signals are decoded or a point is reached where an iteration fails. Though significant progress has been made in MPR techniques at the physical layer, little attention has been paid to the design of support protocols at high layers. As not all composite signals are decodable, it is indispensable to avoid harmful collisions (i.e., when the involved signals cannot be separated). In particular, as there are specific requirements to ensure the feasibility of an MPR method, it is necessary to coordinate the transmissions carefully to meet the requirements. Dealing with interference is one of the primary challenges in wireless communication system design. In the literature, there are two major interference models: the protocol model and the physical model. Though several extensions have been introduced to the models to deal with MPR, the unique feature of SIC is not captured accurately. For example, in [1], the protocol model is extended by increasing the number of permitted interferers from zero to N (N≥1). The extension ignores the constraint in the received signal strength imposed by the sequential detection in SIC. To better understand scheduling performance, here we introduce an layered physical model and a layered protocol model, i.e., M-protocol model, where M is a pre-defined system parameter, to characterize the impact of SIC. We take successive interference cancellation (SIC) [5] as an example of MPR to study scheduling performance in a network with SIC. The protocol design has been considered only recently, e.g., link scheduling [6, 7] and topology control [8], in a network with SIC. However, it is lacking in understanding the generic behavior of a network with SIC. To completely understand the effect of SIC, in this paper, we study the scheduling performance from three different aspects. First, given a scheduling scheme that is unaware of SIC, we analyze the effect of SIC on the approximation performance of the scheme. In our recent work [6, 7], we show that link scheduling with SIC is NP-hard in both the M-protocol model and the layered physical model. As there is no optimal solution with polynomial time complexity for any NP-hard problem, we

2

resort to an approximation scheme to perform the scheduling. We demonstrate that, in both the M-protocol model and the layered physical model, the same order of the approximation ratio is achieved for several SIC-unaware scheduling schemes, no matter whether or not SIC is available. A key insight is that the number of simultaneous transmissions increases at most by a limited factor after SIC is applied. The second contribution is the derivation of the capacity of a network with SIC and the finding that it has the same order as that √ without SIC. In the M-protocol model, the capacity is O( n) where n is the total number of nodes. In the layered physical model, if the transmission power can be set arbitrarily, O(n) capacity is achievable; otherwise, the capacity falls down to O(n(η−1/η) ), where η is the path loss exponent. In comparison with the result in [9], the capacity order is not changed when SIC is applied. As a result, any scheduling scheme can achieve the same order of the approximation ratio in a network without SIC as that with SIC. The third contribution is the study of the impact of SIC from first principles. In both chain and cell network topologies, SIC improves the performance significantly. The optimal throughput with SIC is 20% to 100% higher than that without SIC. However, unless SIC is properly characterized and exploited, any scheduling scheme cannot effectively utilize the new transmission opportunities. Moreover, there is an essential correlation between the scheduling performance and the usage of the transmission opportunities from SIC. Therefore, to accurately measure the performance of a scheduling scheme, in addition to the approximation ratio, new metrics are required to explicitly characterize the SIC capability. All in all, the results indicate the importance of designing an SIC-aware scheduling scheme, and suggest that: first, SIC can significantly improve the network capacity, and characterizing the impact of SIC is indispensable to exploit the new transmission opportunities; second, the approximation ratio is not sufficient to measure the performance of a scheduling scheme when SIC is available. The findings of this work should shed some light on the protocol design in a network of SIC and the impact of other similar MPR techniques on scheduling. The rest of this paper is organized as follows. Section II overviews the related work and Section III describes the system model. Section IV derives the approximation ratio of two scheduling schemes when SIC is available. Section V analyzes the network capacity and Section VI examines the impact of SIC. We conclude the research in Section VII. The proofs are given in Section VIII.

Scheduling packet transmission in a network without SIC has been considered in [13, 14] based on the protocol model, and in [15, 16] based on the physical model. In [17], a scheduling scheme is proposed to achieve a constant approximation ratio in the protocol model. Also, efficient approximation algorithms in the physical model are given in [16] under the assumption that transmitters can either broadcast at full power or not at all, and in [15, 18] by choosing different transmission powers for different transmitter nodes. The capacity of a random network in both the physical and the protocol models is examined in [9]. The capacity of an ad hoc network is studied in [19] under different topologies and traffic patterns. Also, SIC is shown to improve the performance significantly in various wireless networks [20]. To realize the potential of MPR, network protocols must be designed accordingly. There are some studies to support MPR in a centralized network [12] and in a distributed scenario, e.g., distributed MAC [11] and joint routing and scheduling [10]. SIC-aware protocol design in a network with SIC has only recently been considered. For example, link scheduling in a network with SIC is studied in [6, 7] based on both the protocol and physical models. Also, in [8], topology control is examined in a multi-user MIMO network with SIC. III. SYSTEM MODEL Consider a single-channel wireless network of n stationary nodes (i.e., X = {X1 , . . . , Xn }) and N links. A link is denoted by LS l Rl or Ll (1 ≤ l ≤ N) with transmitter node S l ∈ X and receiver node Rl ∈ X, respectively. We also use Xi (1 ≤ i ≤ n) to denote the position of node Xi and |Xi X j | the distance between two nodes Xi and X j . Assume that: • All nodes are located in a planar area; • The signal removal of SIC is perfect; • The network node is homogenous. Each node has an omni-directional antenna, operates in the half duplex mode, transmits with the same transmission power over the common wireless channel, and is not able to transmit multiple packets simultaneously. • The transmission rate is the same for all transmitter nodes, i.e., rate adaptation is disabled. Note that signal removal is challenging in a near-far situation. In practice, likely they will be residual interference after signal removal even without the near-far effect. We here do not consider the effect of residual interference [20] and leave it as a future work. A. Layered protocol model

II. RELATED WORK In the literature, there are two major interference models: the protocol model and the physical model [9]. To deal with the MPR, the protocol model is extended by increasing the number of permitted interferers from zero to N (N≥1) [10], while the physical model is enhanced by allowing reception with a lower SINR threshold [11]. The model used in [12] correlates the reception probability with the number of concurrent transmissions, while neglecting any difference among transmissions.

In the original protocol model, there is one transmission range and one interference range. A transmission from S i to Ri is successful when S i is within the transmission range of Ri and there is no other active transmitter within the interference range of Ri . We propose a layered protocol model, i.e., the M-protocol (M ≥ 1) model. Here, M is a pre-defined system parameter and, without loss of generality, we assume that M is a bounded constant and independent of the network size (i.e., n). Let rk (1 ≤ k ≤ M) denote the kth transmission range, (1 + δk )rk

3

the kth interference range. In general, we assume that r M > r M−1 > . . . > r1 > r0 = 0 and δk > 0 for all 1 ≤ k ≤ M. Definition 1 Link Li is a k-level link if rk−1 < |S i Ri | ≤ rk . A signal from Xi to X j is a k-level (1 ≤ k ≤ M) signal if rk−1 < |Xi X j | ≤ rk . Then a function U is defined as U(Xi , X j ) = k. In particular, U(Xi , X j ) = ∞ when |Xi X j | > r M . Link L j is a correlated link of Li if, U(S i , Ri ) < ∞, k = U(S j , Ri ) < ∞, and |S i Ri | > (1 + δk )rk . When the two links transmit simultaneously, in order to detect its desired signal (i.e., from S i ), Ri should first detect and remove the signal transmitted from S j . For a link L, suppose there are J (J ≤ N − 1) links active simultaneously with L and D (D ≤ J) of them are correlated links of L. Without loss of generality, all the links are ordered with respect to the distance to the receiver of L as L1 , . . . , L J+1 , where LD+1 is the targeting link L. Suppose |S 1 RD+1 | ≤ . . . ≤ |S J+1 RD+1 | and the set of correlated links is {L1 , . . ., LD }. To successfully detect the signal of LD+1 , the required condition is, for any integer x (1 ≤ x ≤ D), we have: u = U(S x , RD+1 ) < ∞, and for every x < y < J + 1, |S y RD+1 | > (1 + δu )ru .

(1)

B. Layered physical model Let N0 denote the noise power, P the transmission power, and Pij = P/|S j Ri |η the received signal power at Ri from S j , where η is the path-loss exponent and usually 2 ≤ η ≤ 6. Link L j is a correlated link of Li if, at node Ri , the signal of L j is sufficiently strong so that it can be detected in the presence of that of Li . Afterwards, the signal of L j is removed to reduce the interference to Li . The required condition is Pij N0 + Pii

≥θ

(2)

where θ specifies the reception SINR threshold. For a link L, suppose there are J (J ≤ N − 1) links active simultaneously with L and D (D ≤ J) of them are correlated links of L. Without loss of generality, all the links are ordered with respect to the distance to the receiver of L as L1 , . . . , L J+1 , where LD+1 is the targeting link L. Suppose |S 1 RD+1 | ≤ . . . ≤ |S J+1 RD+1 | and the set of correlated links is {L1 , . . ., LD }. To successfully detect the signal of LD+1 , the required condition is, PD+1 x N0 +

P

PD+1 j

≥ θ, ∀x ≤ D

N0 +

P

PD+1 j

≥ θ.

(x+1)≤ j≤J+1 PD+1 D+1 (D+2)≤ j≤J+1

(3)

It is clear that the protocol model and the physical model are a special case of the two new models, respectively. The original protocol model is the same as the M-protocol model when M = 1, and the original physical model is the same as the layered physical model when no iterative detection is allowed.

IV. MAINTENANCE OF THE ORDER OPTIMALITY For packet transmission, time is partitioned to slots of a constant duration. Each slot is for transmission of one packet. To measure the performance of a scheduling scheme, schedule length is defined as the total number of time slots used by the scheme. The objective of a scheduling scheme is to allocate each link at least one slot while assuring the schedule length is as short as possible. For a scheduling scheme A, approximation ratio is defined as the ratio of the schedule length of A to the optimal one, which is the minimum number of slots to schedule all the links. Below, for each interference model, we choose a scheduling scheme to examine its performance in a network with SIC. A. Scheduling Based on the M-Protocol Model The scheduling scheme shown in Algorithm 1 is similar to that presented in [6] except the definitions of the incoming and outgoing degrees. We show that, based on the M-protocol model, it achieves a constant approximation ratio no matter whether or not SIC is available. Algorithm 1: Scheduling based on the M-protocol model

1 2 3 4 5 6 7

8

9 10

Data: A set of links located arbitrarily on the plane Result: A feasible schedule S LO U ← all links; repeat Find a link L in U that has the maximum IN difference and let Ln−m+1 denote the mth chosen link; U ← U − {Ln−m+1}; until U == Ø for i=1 to n do Schedule link Li in the first di available slots such that the resulting set of scheduled links is feasible, where di is the number of slots required by Li ; If currently available slots are not sufficient to schedule di slots for Li , add new slots at the end of the schedule S LO and schedule link Li in these slots; end return the schedule S LO ;

We first introduce the concept of IN difference to order the links to be scheduled. Definition 2 For link LS R , a link that can interfere with the reception of LS R is defined as the interfering link of LS R . Based on the M-protocol model, link LS ′ R′ is an interfering link of LS R when |S ′ R| ≤ (1+δk )rk , where k = U(S , R). The incoming degree of LS R is the number of all interfering links. The disk area centering at R with radius (1 + δk )rk is defined as the interference zone of LS R . Definition 3 For link LS R , a link that is interfered by LS R is defined as the interfered link of LS R . Based on the M-protocol model, link LS ′ R′ is an interfered link of LS R when |S R′ | ≤ (1 + δk′ )rk′ , where k′ = U(S ′ , R′ ). The outgoing degree of LS R is the number of all interfered links. Definition 4 The IN difference of a link is the difference between the incoming degree and the outgoing degree. The scheduling scheme is summarized in Algorithm 1, which has two major procedures.

4

Link ordering: The first link that has the maximum IN difference is chosen. While not all links are scheduled, do the following: select the link with the maximum IN difference; and remove the chosen link. The selection process provides a particular ordering of all links. • Slot allocation: Time slots are assigned to each link from the last one to the first. When the demand of a link is larger than one slot, multiple slots are assigned to meet the demand. If currently available slots are not enough, new slots are allocated to schedule the link. Finally, at every time slot, a feasible link set is constructed. In [6], it is shown that, when the demand is one for every link, the schedule length of Algorithm 1 is bounded by O(∆in ), where ∆in is the maximum incoming degree. It can be verified that this is still valid in the M-protocol model. •

Lemma 1 (From [6]) Based on the M-protocol model, the schedule length reported by Algorithm 1 is at most (2∆in + 1). Now we present the approximation ratio of Algorithm 1. Theorem 1 Based on the protocol model, Algorithm 1 has a constant approximation ratio in a network without SIC. The basic approach to prove the theorem is to divide the interference zone of a link into several regions. Fig. 1 shows the partition of the interference zone: first draw K circles within the zone with radius dk = k(1+δ)r (k = 1, . . . , K) and K then divide the area between two consecutive circles into ⌈ 2π α⌉ regions, where α ∈ (0, 2π) is a constant determined by r and δ. The shadow area in Fig. 1 shows an example of the region, which is termed as a (k − 1, k, α) region. The endpoints of the region are denoted by Ak,1 , Ak,2 , Ak−1,1 and Ak−1,2 , where Ak,1 and Ak,2 reside on the kth circle (i.e., with radius k(1+δ)r K ) and Ak−1,1 and Ak−1,2 reside on the (k-1)th circle (i.e., with radius (k−1)(1+δ)r ). Afterwards, it can be shown that (i) the number K of regions, i.e., K⌈ 2π α ⌉, depends on r and δ only, and (ii) the incoming links whose senders are in the same region must interfere with each other. In consequence, at least Ω(∆in ) slots are required for an optimal schedule. Theorem 2 Based on the M-protocol model, Algorithm 1 has a constant approximation ratio in a network with SIC. Theorem 1 is a special case of Theorem 2 when M = 1. The approach to derive the result when M ≥ 2 (i.e., when SIC is available) is similar except three differences: (i) For a link Li , the interference range is (1 + δu )ru , where u = U(S i , Ri ); (ii) The set of incoming links is divided into M subsets such that every link in the kth subset is a k-level link; (iii) The number of regions depends on {r1 , . . . , r M , δ1 , . . . , δ M }. To our best knowledge, Algorithm 1 is the first scheduling scheme that is shown to achieve a constant approximation ratio in a network with SIC. However, though Algorithm 1 may take advantage of some transmission opportunities from SIC, its design is not SIC-aware. For example, the incoming degree of link Li counts all the links interfering Li in the M-protocol model. When L j is a correlated link of Li , though the impact of the L j interference on Li is removable, L j is still included in counting the incoming degree of Li . One may argue that the result likely attributes to the

Fig. 1: Partition of the interference zone of link LS i Ri . The shadow area is called a (k −1, k, α) region with four endpoints Ak,1 , Ak,2 , Ak−1,1 , Ak−1,2 .

simplicity of the interference model, e.g., no accumulative effect of interference is considered. Next, we show a similar behavior in an accumulative interference model.

B. Scheduling Based on the Layered Physical Model We study the performance of the scheduling scheme given in Algorithm 2 [16]. It consist of two steps: first, the problem instance is partitioned into disjoint link length classes; then, a feasible schedule is constructed for each length class using a greedy strategy. For more details, please refer to [16]. For a non-negative integer x, we say that, Li is an x-class link when 2 x ≤ |S i Ri | < 2 x+1 . Algorithm 2: Scheduling based on the physical model [16]

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Data: A set of links located arbitrarily on the plane Result: A feasible schedule Let R = R0 , . . . , Rlog(lmax ) such that Rk is the set of links Li of length 2k ≤ |S i Ri | < 2k+1 ; t = 1; for all Rk , ∅ do Partition the plane into squares of width µ · 2k ; 4-color the cells such that no two adjacent squares have the same color; for j=1 to 4 do Select color j; repeat For each square A of color j, pick one link Li ∈ Rk with receiver Ri in A, assign it to slot t; t = t + 1; until all links of Rk in the selected squares are scheduled; end end return the schedule;

Definition 5 For a link set L, let length diversity, i.e., g(L), denote the number of non-empty length classes. Let N be the

5

k

k

k

k

Fig. 2: (a) Partition of the plane into square grid cells of side µ · 2k ; (b) Partition of a cell into 9 subcells of side µ · 2k /3. set of non-negative integers, then g(L) is given by g(L) = |{m|m ∈ N; ∃Li , L j ∈ L : ⌊log |S i Ri |/|S i R j |⌋ = m}|. For each 1 ≤ k ≤ g(ALS ), where ALS is the set of all links, the plane is partitioned into square grid cells of side µ · 2k , where u = 4(8θ ·

η − 1 1/η ) . η−2

Fig. 2 (a) shows an example of the partition. Let C be the number of cells and Lyx be the set of links Li whose receiver is located in the yth cell and 2 x ≤ |S i Ri | < 2 x+1 . Then we choose a special set Lkm such that |Lkm | = max1≤x≤g(ALS ),1≤y≤C {|Lyx |} and let ∆km = |Lkm |. Lemma 2 (From [16]) Based on the physical model, the schedule length of Algorithm 2 is at most O(g(ALS ) · ∆km ) in a network without SIC. When SIC is available, the performance of the scheduling scheme is stated as follows. Theorem 3 Based on the layered physical model with uniform transmission power, the approximation ratio of Algorithm 2 is O(g(ALS )) in a wireless network with SIC. When SIC is applied, the schedule computed by Algorithm 2 is still feasible. Therefore, with Lemma 2, we need to show that the optimal schedule requires at least Ω(∆km ) slots. We further partition a cell of side µ · 2k into 9 sub-cells of side µ k 3 ·2 , as shown in Fig. 2 (b). Afterwards, we bound the number of links in Lkm such that (i) they transmit simultaneously; (ii) Ri is located in the same sub-cell; and (iii) 2k ≤ |S i Ri | < 2k+1 . The upper bound q depends only on the path loss exponent η and the SINR reception threshold θ. As a result, an optimal schedule requires at least ∆km /(9q) slots. It can be verified that, if the transmission power is nonuniform, the length of the optimal schedule is decreased by a factor at most σ, where σ is the ratio of the maximum transmission power to the minimum one. Therefore, the approximation ratio is still O(g(ALS )) when σ is a small constant. If the transmission power can be set arbitrarily, a higher gain can be expected when SIC is available. At this time, the result of Theorem 3 no longer holds. The joint design

of power control and scheduling is beyond the scope of this paper, and we leave it for future work. It is shown in [16] that the approximation ratio of Algorithm 2 is O(g(ALS )) in a network without SIC. Theorem 3 shows that the same order of approximation ratio is achieved when SIC is available. Note that the scheme is unaware of SIC and does not exploit any transmission opportunity from SIC. On the other hand, it is shown that the capacity is significantly increased when SIC is applied [20]. To explore why a SICunaware scheduling scheme can maintain its order optimality in a network with SIC, we are interested in understanding (i) the impact of SIC on the network capacity and (ii) the scheduling performance in practice when SIC is applied. V. CAPACITY ANALYSIS To explore the generic behavior of the scheduling schemes, we analyze the capacity in a network with SIC. Definition 6 ([9]) The network transports one bit-meter when one bit has been transported a distance of one meter toward its destination. The sum of products of bits and the distances over which they are carried is defined as the transport capacity. To analyze the capacity of a wireless network, we scale the network coverage area and consider that n nodes are located arbitrarily in a disk of area A m2 on the plane. The transmission rate over the channel is W bits per second. Each node can transmit λ bits per second on average and the network transports λnT bits over T seconds. The average distance of a link is B, which implies that a transport capacity of λnB bit-meters per second is achieved. To simplify the analysis, we relax the M-protocol model, i.e., replacing (1) by |S iy Rd | > (1 + δux )|S ix Rd |.

(4)

As |S ix Rd | ≤ rux , the new model is more optimistic in the sense that a feasible link set in the “old” M-protocol model is still feasible in the new model. Theorem 4 Based on the M-protocol model, the transport capacity of a network with SIC is bounded as follows √ √ 8 AW √ (5) λnB ≤ √ n π δ where δ = min{δ1 , . . . , δ M }. In the M-protocol model, the minimum in {δ1 , . . . , δ M } helps us to bound the transport capacity of a network with SIC. Note that {δ1 , . . . , δ M } is determined by the system capability (e.g., the decoding policy) and independent of the network size (e.g., n). As shown in [9], the capacity of a network without SIC is characterized by δ M . Hence, a slightly higher bound can be expected when SIC is available. However, the order of the capacity with SIC is the same as that without SIC. Now turn to the accumulative interference model. At first, if arbitrary transmission power is allowed, O(n) capacity can be achieved. Consider a unique receiver node R at the center with transmitter nodes S 1 , . . . , S n−1 at distances as d1 , . . . , dn−1 to

6  

)*,

    







   ,GIZUX\GR[K









Fig. 3: Cumulative distribution function of the factor (1+D)1/η.

R, respectively, where d1 ≥ d2 ≥ . . . ≥ dn−1 . The transmission power levels (e.g., Pi for node S i , 1 ≤ i ≤ n − 1) are given by η

P1 = θ · d1 · N0

Pi = θ · diη · (N0 +

X

1≤k P12 . Suppose that S 1 and S 2 can transmit to R1 separately, i.e., P11 /N0 ≥ θ and P12 /N0 ≥ θ. Consider the effect of RA on SIC first. With SIC, the two nodes can transmit simultaneously when P11 /(N0 + P12 ) ≥ θ. Otherwise, a harmful collision occurs and no signal can be detected. The relation between S 1 and S 2 is binary: either they can transmit simultaneously, or not. In comparison, with the help of RA, simultaneous transmissions can always be sup-

7

i, j

i, j 1

(a)

(b)

Fig. 4: Illustration of a network consisting of multiple chains. The nodes within the rectangles are affected by the transmission from Xi, j to Xi, j+1 .

ported. To combat the interference, however, the transmission rate should be adjusted accordingly. When P11 /(N0 + P12 ) ≥ θ, as the signal from S 1 can be detected and removed first, there is no need for S 1 and S 2 to change the transmission rate. Otherwise, both S 1 and S 2 must use a lower transmission rate to tolerate the mutual interference. Next, consider the effect of SIC on RA. With SIC, RA can utilize the channel more efficiently. Take S 2 as an example. Without SIC, the chosen transmission rate must tolerate both noise and the interference from S 1 . On the other hand, with the help of SIC, when P11 /(N0 + P12 ) ≥ θ, the signal from S 1 can be removed first so that it is enough to consider the effect of noise only. Eventually, a higher transmission rate can be chosen by S 2 . In summary, to achieve the optimal network performance, RA and SIC should be deployed jointly. It is thus important to extend the analysis to investigate the joint effect of SIC and RA, which is one of our ongoing works. VI. SCHEDULING PERFORMANCE IN PRACTICAL NETWORKS The scheduling performance is examined in a network with SIC from first principles. For simplicity, we limit the discussion to the M-protocol model with M = 2. Note that when M is larger, a higher performance gain can be expected. Hence, the result for M = 2 is a lower bound of the gain from SIC. Let r1 = 35 r2 , δ2 = 1, and δ1 = 1/2. We investigate the scheduling performance in two scenarios. • Chain topology: Each chain contains a sufficiently large number of nodes located on a line. The network comprises one or more chains. • Cell topology: In each cell, there is a receiver node at the center of a circle area and one or more transmitter nodes uniformly located within the area. The network comprises one or more cells. A. Chain Topology Fig. 4 illustrates a network consisting of multiple chains, where rH = r2 and rV = r1 . We assume that the number of nodes in a chain is sufficiently large and denote the node at the ith (i ≥ 1) chain by Xi,1 , Xi,2 , . . .. At the ith (i ≥ 1) chain, every Xi, j ( j ≥ 1) transmits at 1pkt/slot to Xi, j+1 .

Fig. 5: A snapshot of the optimal schedule at one slot in a network without SIC: (a) three chains and (b) four chains.

We first derive the optimal average throughput in a network without SIC. As the transmission distance is r2 , the interference range is 2r2 . Thus, a node can communicate directly with its neighbor nodes. In Fig. 4, the five rectangles cover the nodes that cannot transmit or receive simultaneously with the ongoing transmission (i.e., Xi, j → Xi, j+1 ). One chain: When node X1, j transmits to X1, j+1 , there are two interfered nodes (X1, j−2 and X1, j−1 ) that cannot receive packets from a node other than X1, j , and two interfering nodes (X1, j+2 and X1, j+3 ) that cannot transmit simultaneously. The distance between two active transmitters is at least three hops. Hence, the average optimal throughput is 14 pkt/s. Two chains: The transmission at one chain can affect that at the other. For example, when node X1, j transmits to X1, j+1 , in addition to the four nodes in the first chain (i.e., {X1, j−2 , X1, j−1 , X1, j+2 , X1, j+3 }), there are two interfered nodes (X2, j−1 and X2, j ), and two interfering nodes (X2, j+1 and X2, j+2 ) in the second chain. There is no spatial reuse among any four consecutive nodes in one chain and the four neighbors in the other. The average optimal throughput is 18 pkt/s without SIC. Three chains: The distance between X1, j ( j ≥ 1) and X3, j is 2 ×√35 r2 < 2r2 . The distance between X1, j and X3, j−1 (or X3, j+1 ) is 561 r2 < 2r2 . Hence, there is no spatial reuse among twelve nodes at the three chains (i.e., X1, j to X1, j+3 , X2, j to X2, j+3 and X3, j to X3, j+3 , for j ≥ 1). A snapshot of the optimal schedule at one slot is shown in Fig. 5 (a). The optimal average throughput is 19 pkt/s. Four or more chains: The distance between X1, j ( j ≥ 1) and X4, j is 3 × √35 r2 < 2r2 . The distance between X1, j and X4, j−1 (or X4, j+1 ) is 106 5 r2 > 2r2 . Spatial reuse is feasible between a node at the jth chain and that at the ( j+3)th chain. Thus, an optimal schedule is to schedule the nodes in the first and fourth chains together and the nodes in the second and third chains in separate slots. A snapshot of the optimal schedule at one slot is shown in Fig. 5 (b). At each slot, four packets are transmitted among every 44 nodes. Therefore, the optimal 4 1 average throughput is 44 = 11 pkt/s. When there are more than four chains, note that the transmission at the first chain does not affect that at the fifth chain. The same throughput can be achieved as that in a network with four chains. Now consider the impact of SIC. When there is only one

8

Throughput (pkts/slot)

0.28 0.24 0.2

With SIC Without SIC Approximation

0.16 0.12 0.08 0.04 0

1

2

3

4 5 6 7 Number of chains

8

9

10

Fig. 7: Average throughput versus the number of chains.

Fig. 6: A snapshot of the optimal schedule at one slot in a network with SIC: (a) two chains; (b) three chains; and (c) four chains. (a)

chain, as the distance between any two nodes is at least r2 , SIC cannot be applied. Thus, the optimal average throughput with SIC is the same as that without SIC. Consider a network of two chains. When node X1, j transmits to X1, j+1 , as the distance between X1, j and X2, j is r1 , X2, j can detect the signal from X1, j in the presence of a signal from X2, j−1 . This leads to a new transmission opportunity, i.e., X1, j and X2, j−1 can transmit simultaneously. Finally, a snapshot of the optimal schedule at one slot is shown in Fig. 6 (a). The optimal average throughput is 41 pkt/s. For a network of three or more chains, due to the interference among different chains, a fewer number of simultaneous transmissions can be supported. The optimal schedules for three and four chains are shown in Fig. 6 (b) and (c), respectively. Table I summarizes the optimal average throughput for various network sizes. The SIC gain is defined as (T w −T wo )/T wo, where T wo and T w refer to the optimal average throughput in a network without and with SIC, respectively. When there is no SIC, spatial reuse is possible only after the signal is sufficiently attenuated. Therefore, with an increase of the number of 1 chains, the throughput decreases from 14 to 11 . In comparison, when SIC is available, simultaneously transmission is feasible even when the transmitter nodes are close to each other. Hence, SIC helps to obtain more spatial reuse and a much higher network throughput. The performance gain ranges from 29% to 100%. TABLE I: Throughput comparison in chain topology. Scenario Single chain Two chains Three chains Four or more chains without SIC 1/4 1/8 1/9 1/11 with SIC 1/4 1/4 1/7 1/7 SIC Gain n/a 100% 29% 57%

A scheduling scheme unaware of SIC cannot exploit the transmission opportunity from SIC. For example, the simultaneous transmissions of Xk, j+1 to Xk, j+2 and Xk+1, j to Xk+1, j+1 (k, j ≥ 1) are prohibited without the capability of SIC. Therefore, unless the unique feature of SIC is characterized, any scheduling scheme will fail to recognize the new transmission

(b)

interference

Guard nodes

interference

Fig. 8: A snapshot of the optimal schedule at one slot in a network of four chains when rV = 2r2 /5: (a) with SIC; (b) without SIC.

opportunities. To verify the analysis, we conduct simulation to investigate the performance of Algorithm 1. Fig. 7 compares the throughput of the approximation scheme with the optimal ones. Though the design of the approximation scheme is not SIC-aware, the scheduling scheme can exploit some transmission opportunities from SIC when allocating time slot for a link. As an approximation scheduling scheme, it is naturally sub-optimal. However, the throughput of the scheduling scheme is close to the optimal one without SIC. This means that the sub-optimality is compensated to a large extent by the usage of the SIC capability. Nevertheless, the throughput is much lower than the optimal one with SIC, which indicates that it is challenging to exploit all transmission opportunities from SIC. With different node distances, the degree spatial reuse and the opportunities of simultaneous transmissions are also different. For example, when rV changes from 53 r√2 to 25 r2 , the distance between X1, j and X4, j−1 (or X4, j+1 ) is 561 r2 < 2r2 . As a result, there is interference between the nodes at the jth chain and those at the (j+4)th chain. The optimal schedule in a network of four chains are shown in Fig. 8. In comparison with Fig. 5 (b) and Fig. 6 (c), a column of nodes are required as guard nodes to avoid the interference. The throughput without 1 SIC decreases to 12 and that with SIC is 18 . The gain provided by SIC is 50%. We also perform another group of experiments: both rV and rH are chosen randomly between 12 r1 and r2 ; the number of chains ranges from 4 to 10. Finally, about 120 different



Normalized Throughput

'\KXGMKZNXU[MNV[ZMGOT

9

     











1.9 1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 0

+^VKXOSKTZ 

Fig. 9: Average throughput gain of SIC in the experiments with different chain topologies.

ȡ=6 2

4

ȡ=10

ȡ=14

6 8 10 12 14 16 Number of receiver nodes

18

20

Fig. 11: Normalized throughput with SIC versus the number of receiver nodes with different densities of the transmitter nodes.

4UXSGRO`KJ:NXU[MNV[Z

   

5VZOSGR]OZN9/) 5VZOSGR]OZNU[Z9/) 'VVXU^OSGZOUT

    

Fig. 10: Illustration of a single cell network.

topologies are generated. In addition, for every node Xi, j , a transmission probability (0.6 ∼ 0.8 in the experiments) is adopted to determine whether or not it transmits to Xi, j+1 . Fig. 9 shows the average throughput gain of SIC in different networks. There is no gain provided by SIC in about 20 topologies, where every two links do not satisfy the constraints (i.e., (1)). In the remaining 100 networks, the throughput gain provided by SIC is on average 50% and up to 100%. B. Cell Topology Now we consider the cell topology. A cell is a disk area with radius r2 . In each cell, there is a unique receiver node at the center and several transmitter nodes located uniformly within the cell. Let ρ denote the node density, i.e., the number of the transmitter nodes per unit area. 9 Note that the 1-level interference range is (1 + 12 ) 53 r2 = 10 r2 . Consider a network of a single cell, as shown in Fig. 10. Let E1 denote the area with distance to the receiver no more 9 9 than r1 , E2 between r1 and 10 r2 , and E3 between 10 r2 and r2 . We use E1 , E2 and E3 to denote the three sets of transmitter nodes in the three areas, respectively. With SIC, the nodes in E1 can transmit simultaneously with those in E3 . The optimal schedule length is |E2 | + max{|E1 |, |E3|}. Without SIC, as no concurrent transmissions are permitted, the optimal schedule length is |E1 | + |E2 | + |E3 |. 19π 2 2 9π 2 The areas of E1 , E2 and E3 are 9π 25 r2 , 20 r2 and 100 r2 , respectively. Then, the optimal schedule length with SIC is 36π 2 81π 2 2 ρ( 9π 20 r2 + 100 r2 ) = 100 r2 ρ. In comparison, the optimal schedule length without SIC is πr22 ρ. The performance gain is 19%. When a network comprises two or more cells, the performance depends on the intersection among different cells.







      4[SHKXULXKIKO\KXTUJKY





Fig. 12: Normalized throughput versus the number of receiver nodes in the cell topology.

With SIC, the simultaneous transmissions in a single cell are always feasible in a network of multiple cells. The gain in a single cell is thus a lower bound of that in a larger network. As it is impossible to accurately derive the optimal schedule length in a large network, we use √ simulation to investigate the performance. We set r2 = 1/ π and randomly chooses the positions of the receiver nodes in a 3r2 × 3r2 plane. For each receiver node R, in the area centering at R with radius r2 , the transmitter nodes are generated uniformly with density ρ. Fig. 11 shows the normalized throughput with SIC versus the number of receiver nodes with different ρ. The 95% confidence interval is also shown. The optimal average throughput without SIC is normalized to 1. When the density is larger, a slightly higher throughput is obtained but the difference is not significant. Fig. 12 shows the normalized throughput versus the number of receiver nodes when ρ = 10. The throughput of the scheduling scheme is close to the optimal one without SIC but much lower than that with SIC. As a significant gain as large as 80% is obtained when SIC is available, it is important to explore how to exploit all the new transmission opportunities. Finally, we investigate the correlation between the scheduling performance and the usage of the SIC capability. The simulation settings are the same except that: (i) the network plane is 5r2 × 5r2 , and (ii) the maximum number of receiver nodes is 50 and ρ = 10. Fig. 13 shows the throughput percentage versus the SIC utilization ratio when the number of receiver nodes ranges from 30 to 50. The throughput percentage is defined as the ratio of the throughput of the

Throughput percentage

10

1

VII. CONCLUSIONS AND FUTURE WORK

0.8 0.6 0.4 0.2 0 0

0.05

0.1 0.15 SIC utilization ratio

0.2

Fig. 13: Throughput percentage versus the SIC utilization ratio in the cell topology.

scheduling scheme to the optimal one with SIC. The SIC utilization ratio is defined as the ratio of the number of used correlated links to the total number of correlated links. Let L1 be a correlated link of L2 , L1 is used when the same time slot is assigned to L1 and L2 . The correlation coefficient between the throughput percentage and the SIC utilization ratio is given in Table II with different numbers of receiver nodes. For each number, the experiments are repeated 500 times with different random seeds. It is clear that with a higher utilization ratio, a higher throughput can be expected. This is not surprising since a higher utilization ratio means that a larger number of transmission opportunities from SIC have been exploited. In addition, the fact that the correlation coefficients in Table II are close to or larger than 0.5 indicates the essential correlation between the scheduling performance and the usage of the SIC capability. The relatively low utilization ratio points out that there is still a large room for future work to fully exploit the SIC capability in the design of protocols such as a link scheduling scheme.

TABLE II: Correlation coefficient between the throughput percentage and the SIC utilization ratio. Number of receiver nodes 10 20 30 40 50 Correlation coefficient 0.537 0.508 0.485 0.552 0.488

This paper investigates the scheduling performance in wireless networks with successive interference cancellation. After introducing two interference models to capture the impact of SIC, we show that the capacity in a network with SIC has the same order as that without SIC. It is therefore not surprising that a scheduling scheme unaware of SIC maintains its order optimality when SIC is available. We examine the impact of SIC from first principles and find out that a significant throughput gain between 20% and 100% is obtained from SIC. Moreover, the performance gain of a scheduling scheme is essentially correlated with the usage of the transmission opportunities from SIC. This work demonstrates the importance of designing an SIC-aware scheduling scheme, and suggests that the approximation ratio is not a sufficient indicator of the scheduling performance when SIC is available. There are several directions to extend the work. First, it is one of our ongoing works to define a performance metric to properly evaluate the usage of the SIC capability for a scheduling scheme. Second, it is important to consider the joint design of link scheduling and power control in a network with SIC. Third, it is necessary to consider the effect of imperfect signal removal, especially in a near-far situation. Finally, it is interesting to study link scheduling in a network with both SIC and rata adaptation. Currently, we do not consider the effect of rate adaptation and the present studies should be revisited carefully when the transmission rate is adjusted adaptively. VIII. SUMMARY OF THE PROOFS Proof of Theorem 1: With Lemma 1, it is sufficient to show that the optimal schedule length is at least Ω(∆in ). Suppose the incoming degree of LS R is ∆in . For any incoming link LS ′ R′ of LS R , we have |S ′ R| ≤ (1 + δ)r. Using R as the center, we can draw K circles with radius dk = (1 + δ)kr/K (k = 1, . . . , K) and divide the interference zone of LS R into several regions (cf. Fig. 1). The number of the regions is at most K⌈2π/α⌉. Both K and α are constants determined by r and δ. The values of them will be specified later. For a (k − 1, k, α) region, letting D(k − 1, k, α) be the maximum distance between any two points in the region, we have D(k − 1, k, α) = max{|Ak,1 Ak−1,1 |, |Ak,1 Ak−1,2 |, |Ak,1 Ak,2 |}.

As a common metric for all approximation algorithms, approximation ratio fails to carry sufficient information about the usage of the transmission opportunities from SIC. As a result, to accurately measure the performance of a scheduling scheme, in addition to the approximation ratio, new metrics are required to explicitly characterize the effect of SIC. In summary, two important observations are obtained. First, though there is no improvement in the capacity order, the performance gain obtained from the new transmission opportunities due to SIC is significant, i.e., between 20% and 100%. Second, the approximation ratio is not a sufficient indicator of the scheduling performance in a network with SIC. Even for a scheduling scheme with a constant approximation ratio, there are still many transmission opportunities not yet exploited.

It is clear that |Ak,1 Ak−1,1 | = (1 + δ)r/K, and q 2 − 2d d |Ak,1 Ak−1,2 | = dk2 + dk−1 k k−1 cos α √ |Ak,1 Ak,2 | = dk 2 − 2 cos α. Now, we show that, if D(k − 1, k, α) ≤ δ · r

(7)

then any two incoming links whose senders are in the same region must interfere with each other. Considering two such incoming links, e.g., LS 1 R1 and LS 2 R2 , then |S 1 R2 | ≤ |S 1 S 2 | + |S 2 R2 | ≤ r + D(k − 1, k, α) ≤ (1 + δ)r.

11

(9)

where Z1 ≤ K1 ⌈2π/α1 ⌉. Then S1 is divided into Z1 subsets such that the links whose senders are in the same region are grouped together. Consider a (k − 1, k, α1 ) region, and let D(k − 1, k, α1 ) be the maximum distance between any two points within a (k − 1, k, α1 ) region. If

(10)

D(k − 1, k, α1 ) ≤ δ1 · r1

Next, we show how to choose K and α to satisfy (7), which is equivalent to the following three inequalities (1 + δ)r/K ≤ δ · r q 2 − 2d d dk2 + dk−1 k k−1 cos α ≤ δ · r √ dk 2 − 2 cos α ≤ δ · r.

(8)

then any two incoming links in S1 whose senders are in the same region must interfere with each other. For two such incoming links, e.g., LS 1 R1 and LS 2 R2 , we have

Substituting dk−1 = dk − (1 + δ)r/K in (9), we have f (dk ) =2(1 − cos α) · dk2 − 2(1 − cos α) ·

1+δ (1 + δ)2 2 · r · dk + ·r K K2

(11)

≤ δ2 · r 2 . It is clear that f (dk ) monotonically increases with dk when dk > (1 + δ)r/K (which is true when k > 1). Thus, f (dk ) is maximized when dk = (1 + δ)r (i.e., k = K). Substituting dk = (1 + δ)r in (11) and rearranging, we obtain (1 − xA)2 + 1 − A2 ≤ cos α 2(1 − xA)

(12)

where A = δ/(1 + δ) and x = 1/(K · A). Choosing K = (1 + δ)2 /δ2 and substituting it in (12) yield cos α ≥ 1 −

(13)

δ2 . 2(1 + δ)2

It can be verified that, when K = (1 + δ)2 /δ2 and α = δ2 arccos(1 − 2(1+δ) 2 ), (8) - (10) are all satisfied. Now we divide the set of the incoming links of LS R into several subsets such that the incoming links whose senders are in the same region are grouped together. The number of the subsets is at most K⌈ 2π α ⌉. As the links in the same group must interfere with each other, the least number of slot required by an optimal solution is ∆in = Ω(∆in ). K⌈2π/α⌉ Proof of Theorem 2: With Lemma 1, it is sufficient to show that the optimal schedule length is at least Ω(∆in ). Suppose the incoming degree of LS R is ∆in . First, consider the case of M = 2. Let S1 denote the set of the incoming links LS ′ R′ of LS R such that |S ′ R′ | ≤ r1 and S2 the set of the remaining incoming links. Obviously, |S1 | + |S2 | = ∆in . Then, 2π • We can divide S1 into at most K1 ⌈ α ⌉ subsets, where K1 1 and α1 are determined by {r1 , r2 , δ1 , δ2 }. The links in the same subset interfere with each other. 2π • We can divide S2 into at most K2 ⌈ α ⌉ subsets, where K2 2 and α2 are determined by {r1 , r2 , δ1 , δ2 }. In each subset, at most two links can transmit simultaneously. The optimal schedule length is at least

|S 1 R2 | ≤ |S 1 S 2 | + |S 2 R2 | ≤ r1 + D(k − 1, k, α1 ) ≤ (1 + δ1 )r1 . With the same process as in the proof of Theorem 1, one can obtain a feasible setting of K1 and α1 to satisfy (13), i.e., r K1 = β2 , and α1 = arccos(1 − 2β1 2 ), where β = (1+δ) δ1 · r1 . As U(S , R) can be 1 or 2, we set β = max{

(1 + δ1 ) (1 + δ2 ) r2 , · }. δ1 δ1 r1

For S2 , we can draw K2 circles centering at R with radius dk = (1 + δ)r/K2 (k = 1, . . . , K2 ). Similarly, we divide the interference zone into Z2 regions, where Z2 ≤ K2 ⌈2π/α2⌉, and S2 into Z2 subsets such that the links whose senders are in the same region are grouped together. Consider a (k − 1, k, α2 ) region, and let D(k − 1, k, α2 ) be the maximum distance between any two points within a (k − 1, k, α2 ) region. If D(k − 1, k, α2 ) ≤ δ2 · r2

(14)

then at most two incoming links in S2 whose senders are in the same region can transmit simultaneously. For two such incoming links, e.g., LS 1 R1 and LS 2 R2 , we have |S 1 R2 | ≤ |S 2 R2 | + |S 1 S 2 | ≤ r2 + D(k − 1, k, α2 ) ≤ (1 + δ2 )r2 . As M = 2, in a composite signal, at most one signal can be removed by SIC. Hence, when three or more links transmit simultaneously, at any receiver node, at least one interfering signal that can interfere the two-level signal is retained. As every link in S2 is a two-level link, a detection failure must occur when three or more links in the same subset of S2 transmit simultaneously. With the same process as in the proof of Theorem 1, a feasible setting of K2 and α2 can be derived to satisfy (14): K2 = β′2 , and α2 = arccos(1 − 2β1′2 ), where β′ = max{

(1 + δ2 ) (1 + δ1 ) r1 , · }. δ2 δ2 r2

Now turn to the case of M ≥ 3. We divide the incoming links of LS R into M groups: for 1 ≤ j ≤ M, S j contains the incoming link LS ′ R′ such that r j−1 < |S ′ R′ | ≤ r j . Then, |S1 | |S2 | ∆in ∆in max{ , } ≥ min{ , } • We can divide S1 into K1 ⌈ 2π α1 ⌉ subsets, where K1 and α1 K1 ⌈2π/α1 ⌉ 2K2 ⌈2π/α2 ⌉ 2K1 ⌈2π/α1⌉ 4K2 ⌈2π/α2 ⌉ are determined by {r , . . . , r M , δ1 , . . . , δ M }. The links in 1 ∆in in the same subset interfere with each other. = = Ω(∆ ). 2π max{2K1 ⌈2π/α1 ⌉, 4K2 ⌈2π/α2 ⌉} • We can divide S j (2 ≤ j ≤ M) into K j ⌈ α ⌉ subsets, where j Let r = rU(S ,R) , δ = δU(S ,R) . For S1 , we can draw K1 circles K j and α j are determined by {r1 , . . . , r M , δ1 , . . . , δ M }. In centering at R with radius dk = (1 + δ)r/K1 (k = 1, . . . , K1 ) each subset, at most two links can transmit simultaneand then divide the interference zone of LS R into Z1 regions, ously.

12

The optimal schedule length is at least |S1 | |S2 | |S M | , ,..., } K1 ⌈2π/α1 ⌉ 2K2 ⌈2π/α2 ⌉ 2K M ⌈2π/α M ⌉ ∆in ∆in ∆in , . . . , } , ≥ min{ 2MK2 ⌈2π/α2 ⌉ 2MK M ⌈2π/α M ⌉ MK1 ⌈ 2π α ⌉

max{

1

∆in M · max{K1 ⌈2π/α1 ⌉, 2K2 ⌈2π/α2 ⌉, . . . , 2K M ⌈2π/α M ⌉} = Ω(∆in ). =

First, for 1 ≤ i ≤ M, 3 ≤ j ≤ M, define

(1 + δ1 ) r1 (1 + δ M ) r M · ,..., · } δi ri δi ri (1 + δ1 )r1 (1 + δ M )r M = max{ ,..., } r j−1 − (1 + ξ)r j−2 r j−1 − (1 + ξ)r j−2

βi = max{ β j(2)

where ξ ∈ (0, min{ rr12 , . . . , rrM−1 } − 1) is a small constant. M−2 Let r = rU(S ,R) , δ = δU(S ,R) . For each 1 ≤ j ≤ M, we can draw K j circles centering at R with radius dk = (1+δ)r/K j (k = 1, . . . , K j ), and divide the interference zone into Z j regions, where Z j ≤ K j ⌈2π/α j ⌉, and S j into Z j subsets such that the links whose senders are in the same region are grouped together. The processes of both S1 and S2 are similar to that when M = 2. A feasible setting for S1 is K1 = 1/β21 , and α1 = arccos(1 − 2β1 2 ), while a feasible setting for S2 is K2 = 1/β22 , 1

and α2 = arccos(1 − 2β1 2 ). 2 For S j (3 ≤ j ≤ M), consider a (k − 1, k, α j ) region, and let D(k − 1, k, α j ) be the maximum distance between any two points within a (k − 1, k, α j ) region. If D(k − 1, k, α j ) ≤ δ j · r j

(15)

D(k − 1, k, α j ) ≤ r j−1 − (1 + ξ)r j−2

(16)

and

then at most two incoming links in S j whose senders are in the same region can transmit simultaneously. For two such incoming links, e.g., LS 1 R1 and LS 2 R2 , we have |S 1 R2 | ≤ |S 2 R2 | + |S 1 S 2 | ≤ r j + D(k − 1, k, α j ) ≤ (1 + δ j )r j . On the other hand, |S 1 R2 | ≥ |S 2 R2 | − |S 1 S 2 | ≥ r j−1 − D(k − 1, k, α j ) > r j−2 . Thus, at R2 , the strongest interfering signal is at most a ( j − 1)-level signal, and the weakest one can at least interfere the j-level signal. As every link in S j is a j-level link, any three links in S j whose senders are in the same region cannot transmit simultaneously. With the same process as in the proof of Theorem 1, to satisfy (15), a feasible setting can be derived as K j(1) = 1/β2j , and α j = arccos(1 − 2β1 2 ). Similarly, to satisfy (16), a feasible j

setting is K j(2) = 1/β2j(2) , and α j(2) = arccos(1 − 2β12 ). j(2) Finally, when 3 ≤ j ≤ M, a feasible setting is given by K j = max{K j(1) , K j(2) } and α j = min{α j(1) , α j(2) }. Proof of Theorem 3: With Lemma 2, it is sufficient to show that the optimal schedule length is at least Ω(∆km ).

We divide a cell into 9 sub-cells of side µ3 · 2k . For two links Li , L x ∈ Lkm , when Ri and R x are √in the same sub-cell, we have 2k ≤ |S x R x | < 2k+1 and |R x Ri | ≤ 32µ · 2k . Then, √ √ 2µ k 2µ k+1 |S x Ri | ≤ |S x R x | + |R x Ri | ≤ 2 + ·2 =( + 2) · 2k 3 3 and |S x Ri | ≥ max{|R x Ri | − |S x R x |, |S xR x | − |R x Ri |} √ √ 2µ k 2µ ≥| · 2 − 2k+1 | = | − 2|2k . 3 3 Now we bound the number of Li links such that (i) they transmit simultaneously, (ii) Ri is located in the same sub-cell, and (iii) 2k ≤ |S i Ri | < 2k+1 . To ensure the successful detection of LS i Ri , all stronger interfering signals must be removed and the aggregate interference of the remaining should be tolerable. Consider a link LS x R x with |S x Ri | ≤ |S i Ri |, we have θ≤ ≤

P |S x Ri |η

P √ | 2µ/3−2|η 2ηk

≤

P P N0 + |S y Ri |≥|S x Ri | |S yPRi |η |S y Ri |≥|S x Ri | |S y Ri |η √ P √ ( 2µ/3 + 2)η | 2µ/3−2|η 2ηk = . √ P q1 · ( √2µ/3+2) q1 | 2µ/3 − 2|η η ·2ηk N0 +

P

√

)η . On the other hand, Thus, q1 ≤ 1θ ( | √2µ/3+2 2µ/3−2| θ≤ ≤

≤

P 2kη

P N0 + |S y Ri |≥|S i Ri | |S yPRi |η |S y Ri |≥|S i Ri | |S y Ri |η √ P ·( 2µ/3 + 2)η 2kη = . P q2 √ ( 2µ/3+2)η ·2ηk

N0 + q2 ·

P |S i Ri |η

P

P

√ Thus, q2 ≤ ( 2µ/3 + 2)η /θ. Therefore, if there are more than (q1 + q2 ) links satisfying constraints (i)-(iii), for an active link Li , either the signal of Li or that of a correlated link of Li cannot be detected. The least number of time slots required to schedule the links in ∆km is ∆km = Ω(∆km ). 9 · (q1 + q2 )

Proof of Theorem 4: The proof is mainly based on [9]. Considering the bth bit (1 ≤ b ≤ λnT ), suppose that it moves from its origin to its destination in a sequence of h(b) hops, where the hth hop traverses a distance of rbh . Then, h(b) λnT X X b=1 h=1

rbh ≥ λnT B.

(17)

Note that in any slot at most n/2 nodes can transmit. Hence, H :=

λnT X b=1

h(b) ≤

WT n . 2

(18)

Consider a k-level link LS R and a k′ -level link LS ′ R′ (1 ≤ k, k′ ≤ M). If they can transmit simultaneously, we have |RR′ | ≥ |S R′ | − |S R| ≥ (1 + δk′ )|S ′ R′ | − |S R|. |RR′ | ≥ |S ′ R| − |S ′ R′ | ≥ (1 + δk )|S R| − |S ′ R′ |.

13

(2)

(2)

(1)

(1) i

i 2

1

1 2

Fig. 14: Illustration of√the common area between Λ1 and Λ2 : √ (a) r(2) < 2r(1) ; (b) 2r(1) ≤ r(2) < 2r(1) .

of them are correlated links of L. Without loss of generality, all the links are ordered with respect to the distance to the receiver of L as L1 , . . . , L J+1 , where LD+1 is the targeting link L. Suppose |S 1 RD+1 | ≤ . . . ≤ |S J+1 RD+1 | and the set of correlated links is {L1 , . . ., LD }. We have X PD+1 ≥θ· PD+1 x D D+1≤x≤J+1

PD+1 D−1

≥ θ · (PD+1 + D

X

PD+1 x ) ≥ θ(1 + θ)

X

PD+1 x .

D+1≤x≤J+1

X

PD+1 x

D+1≤x≤J

... Adding the two inequalities, we obtain 1 δ (δk′ |S ′ R′ | + δk |S R|) ≥ (|S ′ R′ | + |S R|) 2 2 where δ = min{δ1 , . . . , δ M }. Hence disks of radius 2δ times the lengths of hops centered at the receivers are essentially disjoint. Let r(2) = 2δ rbh and √ r(1) = A/π, then r(2) < 2r(1) . Let Λ1 denote the network coverage area and Λ2 the disk centering at a receiver node Ri with radius r(2) . The common area between Λ1 and Λ2 is minimized when Ri is near the periphery of Λ1 . The shadow regions in Fig. √14 illustrate the common area, for (a) r(2) < √ 2r(1) and (b) 2r(1) ≤ r(2) < 2r(1) . It can be verified that the common area is at least a quarter of Λ2 . As at most Wτ bits can be carried in a slot of length τ from a transmitter to a receiver, at any tth slot (t ≥ 1), we have

P1D+1 ≥ θ(1 + θ)D−1

|RR′ | ≥

h(b) λnT X X b=1

π( δ rh )2 D(b, h, t) 2 b ≤ AWτ 4 h=1

This can be rewritten as h(b) λnT X X 1 b=1 h=1

H

(rbh )2 ≤

16AWT . πδ2 H

(19)

(20)

As the quadratic function is convex, we have h(b) λnT X λnT h(b) X 1 h 2 XX 1 h 2 ( (rb )) ≤ (r ) H H b b=1 h=1 b=1 h=1

which leads to r h(b) λnT X X 1 h 16AWT (rb )) ≤ . H πδ2 H b=1 h=1

(23) ≥ (J − D + 1) · Note that P1D+1 ≤ P and D+1≤x≤J+1 PD+1 x √ P √ . Substituting the last inequality in (23), we obtain (2 A/ π)η P

P ≥ θ(1 + θ)D−1 · (J − D + 1) · As J − D + 1 ≥ 1, we obtain D≤1+

P

√ . (2 A/ π)η √

(24)

√

η log 2√πA − log θ log(1 + θ)

.

(25)

For the detection of LD+1 , we have N0 +

P/|S D+1 RD+1 |η ≥ θ. η (D+1)≤ j≤J+1, j,k P/|S j RD+1 |

P

Rewriting the inequality, we obtain

where D(b, h, t) is one when the bth bit is transmitted over the hth hop at the tth slot, and zero otherwise. As T comprises one or more slots, summing over the slots gives h(b) λnT X X πδ2 h 2 (r ) ≤ WT. 16A b b=1 h=1

D+1≤x≤J+1

(21)

(22)

Substituting (17) and (18) in (22) gives r √ 8 AW √ λnB ≤ n. π δ Proof of Theorem 5: The proof is similar to that Theorem 4 except the difference stemming from the need replace (19) by a new expression. Consider the reception of link L, suppose there are (J ≤ N − 1) links active simultaneously with L and D (D ≤

J J)

(26)

1+θ P P θ N0 + (D+1)≤ j≤J+1 P/|S j RD+1 |η 1+θ P ≤ π η/2 θ N0 + ( 4A ) (J − D + 1)P

(27)

N0 + Hence,

|S D+1 RD+1 |η ≤

which leads to X 1+θ (J + 1)P |S k Rk |η ≤ π η/2 θ N0 + ( 4A ) (J − D + 1)P 1≤k≤J+1 1 + θ 4A η/2 J+1 ·( ) · θ π J−D+1 1 + θ 4A η/2 ≤ · ( ) · (1 + D). θ π As a result, we have ≤

h(b) λnT X X b=1 h=1

of to

P/|S D+1 RD+1 |η θ ≥ . η 1+θ (D+1)≤ j≤J+1 P/|S j RD+1 |

P

(rbh )η ≤

1+θ η · 2 · (A/π)−η/2 · (1 + D)WT. θ

(28)

(29)

The rest of the proof proceeds along lines similar to those of Theorem 4, invoking the convexity of rη instead of r2 . Finally, we obtain √ 2 + 2θ 1/η AW (30) λnB ≤ ( ) · √ · (1 + D)1/η · n(η−1)/η . θ π For non-uniform transmission power, let σ = Pmax /Pmin , where Pmax is the maximum transmission power and Pmin the

14

minimum one. The above statement is still valid when σ is a small constant except a new expression for the upper bound of D. It is easy to see that (23) still holds. Afterwards, we have P P1D+1 ≤ Pmax and D+1≤x≤J+1 PD+1 ≥ (J − D + 1) · (2 √PA/min√π)η . x Then, we obtain Pmax ≥ θ(1 + θ)D−1 · (J − D + 1) ·

Pmin √ √ . (2 A/ π)η

As J − D + 1 ≥ 1, we have σ ≥ Pmax /Pmin ≥ θ(1 + θ)D−1 ·

1

√ . (2 A/ π)η √

(31)

(32)

Rewriting the inequality, we obtain √

D≤1+

η log 2√πA + log σ − log θ log(1 + θ)

.

(33)

Following the same process, we can obtain the capacity finally with the same expression as (30). IX. ACKNOWLEDGEMENTS The authors would like to thank the editor and the reviewers for their valuable comments and suggestions. This work is supported by PCSIRT Program (No.IRT1012), Program for Sci&Tech Innovative Research Team in Higher Educational Institutions of Hunan Province: ”network technology”, NSFC of Hunan Province (11JJ7003), and NSFC (61070203). R [1] J. J. Garcia-Luna-Aceves, H. R. Sadjadpour, and Z. Wang, “Challenges: towards truly scalable ad hoc networks,” in Proc. ACM MOBICOM’07, pp. 207–214, 2007. [2] H. R. Sadjadpour, Z. Wang, and J. J. Garcia-Luna-Aceves, “The capacity of wireless ad hoc networks with multipacket reception,” IEEE Trans. Communications, vol. 58, no. 2, pp. 600–610, 2010. [3] Y. J. Zhang, P. X. Zheng, and S. C. Liew, “How does multiple-packet reception capability scale the performance of wireless local area networks?” IEEE Trans. Mobile Computing, vol. 8, no. 7, pp. 923–935, 2009. [4] M.-F. Guo, X. Wang, and M.-Y. Wu, “On the capacity of kappa-mpr wireless networks,” IEEE Trans. Wireless Communications, vol. 8, no. 7, pp. 3878–3886, 2009. [5] J. Andrews, “Interference cancellation for cellular systems: a contemporary overview,” IEEE Wireless Communications, vol. 12, no. 2, pp. 19–29, 2005. [6] S. Lv, W. Zhuang, X. Wang, and X. Zhou, “Scheduling in wireless ad hoc networks with successive interference cancellation,” in Proc. IEEE INFOCOM’11, pp. 1282– 1290, 2011. [7] ——, “Context-aware scheduling in wireless networks with successive interference cancellation,” in Proc. IEEE ICC’11, pp. 1–6, June, 2011. [8] E. Gelal, K. Pelechrinis, T.-S. Kim, I. Broustis, S. V. Krishnamurthy, and B. Rao, “Topology control for effective interference cancellation in multi-user MIMO networks,” in Proc. IEEE INFOCOM’10, pp. 2357–2365, 2010.

[9] P. Gupta and P. R. Kumar, “The capacity of wireless networks,” IEEE Trans. Information Theory, vol. 46, no. 2, pp. 388–404, 2000. [10] X. Wang and J. J. Garcia-Luna-Aceves, “Embracing interference in ad hoc networks using joint routing and scheduling with multiple packet reception,” in Proc. IEEE INFOCOM’08, pp. 843–851, 2008. [11] G. D. Celik, G. Zussman, W. F. Khan, and E. Modiano, “MAC for networks with multipacket reception capability and spatially distributed nodes,” in Proc. IEEE INFOCOM’08, pp. 1436–1444, 2008. [12] Q. Zhao and L. Tong, “A dynamic queue protocol for multiaccess wireless networks with multipacket reception,” IEEE Trans. Wireless Communications, vol. 3, no. 6, pp. 2221–2231, 2004. [13] T. Nieberg, J. Hurink, and W. Kern, “Approximation schemes for wireless networks,” ACM Trans. Algorithms, vol. 4, no. 4, pp. 1–17, 2008. [14] M. Dinitz, “Distributed algorithms for approximating wireless network capacity,” in Proc. IEEE INFOCOM’10, pp. 1397–1405, 2010. [15] M. Andrews and M. Dinitz, “Maximizing capacity in arbitrary wireless networks in the SINR model: Complexity and game theory,” in Proc. IEEE INFOCOM’09, pp. 1332–1340, 2009. [16] O. Goussevskaia, Y. A. Oswald, and R. Wattenhofer, “Complexity in geometric SINR,” in Proc. ACM MobiHoc’07, pp. 100–109, 2007. [17] W. Wang, X.-Y. Li, O. Frieder, Y. Wang, and W.-Z. Song, “Efficient interference-aware TDMA link scheduling for static wireless networks,” in Proc. ACM MOBICOM’06, pp. 262–273, 2006. [18] O. Goussevskaia, R. Wattenhofer, M. M. Halld´orsson, and E. Welzl, “Capacity of arbitrary wireless networks,” in Proc. IEEE INFOCOM’09, pp. 1872–1880, 2009. [19] J. Li, C. Blake, D. S. J. D. Couto, H. I. Lee, and R. Morris, “Capacity of ad hoc wireless networks,” in Proc. ACM MOBICOM’01, pp. 61–69, 2001. [20] S. P. Weber, J. G. Andrews, X. Yang, and G. de Veciana, “Transmission capacity of wireless ad hoc networks with successive interference cancellation,” IEEE Trans. Information Theory, vol. 53, no. 8, pp. 2799–2814, 2007. [21] M. Franceschetti, M. D. Migliore, and P. Minero, “The capacity of wireless networks: Information-theoretic and physical limits,” IEEE Trans. Information Theory, vol. 55, no. 8, pp. 3413–3424, 2009. ¨ ur, R. Johari, D. N. C. Tse, and O. L´evˆeque, [22] A. Ozg¨ “Information-theoretic operating regimes of large wireless networks,” IEEE Trans. Information Theory, vol. 56, no. 1, pp. 427–437, 2010. [23] S. Lv, X. Wang, and X. Zhou, “On the rate adaptation for IEEE 802.11 wireless networks,” Computer Networks, vol. 54, no. 17, pp. 3173–3186, 2010. [24] S. Sen, N. Santhapuri, R. R. Choudhury, and S. Nelakuditi, “Successive interference cancellation: A back-of-theenvelope perspective,” in Proc. ACM HotNets-IX, pp. 11– 15, 2010.